Abstract

Energy networks face challenges in managing and securing the vast data generated by smart grids. Federated Learning (FL) offers a cost-effective, privacy-aware solution for model training, addressing customer privacy and data breach concerns. However, FL is susceptible to adversarial attacks, particularly data poisoning, which can degrade model accuracy. This study introduces a novel data poisoning attack and a mitigation framework for resource-constrained smart grids. We propose the Centroid Based Anomaly Aware Federated Averaging (CBAA-FedAvg) framework, which achieves a Mean Absolute Percentage Error (MAPE) of 2.7%, closely matching baseline performance. CBAA-FedAvg is a lightweight, sustainable solution that minimizes resource consumption through parameter quantization from 32-bit floating point to 8-bit fixed point and dynamic clustering to reduce computational complexity. Additionally, an automatic stopping criterion is employed to optimize convergence, saving energy and time. The framework demonstrates remarkable resilience against data and model adversarial attacks, offering enhanced security and efficiency compared to state-of-the-art alternatives.